fix(klu): keep topology indices in numpy so nested sax.circuit traces cleanly

src/sax/backends/klu.py

@@ -7,6 +7,7 @@
 import jax
 import jax.numpy as jnp
 import klujax
+import numpy as np
 from natsort import natsorted
 
 import sax
@@ -55,13 +56,46 @@ def analyze_instances_klu(
     model_names = set()
     for i in instances.values():
         model_names.add(i["component"])
-    dummy_models = {k: sax.scoo(models[k]()) for k in model_names}
+    # Build the per-model SCoo with the topology indices ``Si``/``Sj``
+    # cast to numpy. Indices are pure netlist topology (which (i, j)
+    # entries of the model's S-matrix are non-zero) and never depend on
+    # traced parameters, so concretizing them here avoids tracer-only
+    # operations downstream — specifically, the boolean fancy indexing
+    # in ``analyze_circuit_klu``. Only ``Sx`` (the actual S-values) stays
+    # in jnp, where it must be traceable. This is a klu-backend-local
+    # workaround — ``sax.scoo`` and the rest of SAX continue to return
+    # jnp indices for backward compatibility with other backends.
+    dummy_models = {k: _scoo_with_numpy_indices(models[k]()) for k in model_names}
     dummy_instances = {}
     for k, i in instances.items():
         dummy_instances[k] = dummy_models[i["component"]]
     return dummy_instances
 
 
+def _scoo_with_numpy_indices(s: sax.SType) -> sax.SCoo:
+    """Like ``sax.scoo`` but with topology indices forced to numpy.
+
+    For ``SDict`` inputs (the common case) the (Si, Sj) pair is built
+    directly from the dict's keys, so the function is JAX-trace-safe even
+    when the model body has been traced. For ``SCoo`` / ``SDense`` inputs
+    the indices come from ``sax.scoo`` and are eagerly evaluated; if those
+    were produced inside an outer trace this still won't help, but in
+    practice models return ``SDict``.
+    """
+    if isinstance(s, dict):
+        all_ports: dict[str, None] = {}
+        for p1, p2 in s:
+            all_ports.setdefault(p1, None)
+            all_ports.setdefault(p2, None)
+        ports_map = {p: int(i) for i, p in enumerate(all_ports)}
+        Si = np.array([ports_map[p] for _, p in s], dtype=np.int32)
+        Sj = np.array([ports_map[p] for p, _ in s], dtype=np.int32)
+        Sx = jnp.stack(jnp.broadcast_arrays(*s.values()), -1)
+        return Si, Sj, Sx, ports_map
+    si, sj, sx, ports_map = sax.scoo(s)
+    return np.asarray(si, dtype=np.int32), np.asarray(sj, dtype=np.int32), sx, ports_map
+
+
 def analyze_circuit_klu(
     analyzed_instances: dict[sax.InstanceName, sax.SCoo],
     nets: sax.Nets,
@@ -110,8 +144,13 @@ def analyze_circuit_klu(
     n_col = idx
     n_rhs = len(port_map)
 
-    Si = jnp.concatenate(Si, -1)
-    Sj = jnp.concatenate(Sj, -1)
+    # Keep Si / Sj as numpy — they're pure topology (S-matrix nonzero
+    # coordinates per instance) and feed into concrete-only operations
+    # below (boolean fancy indexing, comparisons). Going through jnp here
+    # would turn them into tracers when ``analyze_circuit_klu`` runs
+    # inside an outer JAX trace.
+    Si = np.concatenate(Si, -1)
+    Sj = np.concatenate(Sj, -1)
 
     pairs: set[tuple[int, int]] = set()
     for net in nets:
@@ -120,8 +159,8 @@ def analyze_circuit_klu(
         pairs.add((p1_idx, p2_idx))
         pairs.add((p2_idx, p1_idx))
     sorted_pairs = sorted(pairs)
-    Ci = jnp.array([p[0] for p in sorted_pairs], dtype=jnp.int32)
-    Cj = jnp.array([p[1] for p in sorted_pairs], dtype=jnp.int32)
+    Ci = np.array([p[0] for p in sorted_pairs], dtype=np.int32)
+    Cj = np.array([p[1] for p in sorted_pairs], dtype=np.int32)
 
     Cextmap = {
         int(instance_ports[k]): int(port_map[v]) for k, v in inverse_ports.items()
@@ -130,15 +169,17 @@ def analyze_circuit_klu(
     Cextj = jnp.stack(list(Cextmap.values()), 0)
     Cext = jnp.zeros((n_col, n_rhs), dtype=complex).at[Cexti, Cextj].set(1.0)
 
+    # All in numpy — pure topology, no traced parameters touch this.
     match_2d = Cj[None, :] == Si[:, None]  # (len_Si, len_Cj)
-    CSi = jnp.broadcast_to(Ci[None, :], match_2d.shape)[match_2d]
-    s_idx_grid = jnp.broadcast_to(jnp.arange(len(Si))[:, None], match_2d.shape)
-    cs_s_indices = s_idx_grid[match_2d]
-    CSj = Sj[cs_s_indices]
-
-    Ii = Ij = jnp.arange(n_col)
-    I_CSi = jnp.asarray(jnp.concatenate([CSi, Ii], -1), dtype=jnp.int32)
-    I_CSj = jnp.asarray(jnp.concatenate([CSj, Ij], -1), dtype=jnp.int32)
+    CSi = np.broadcast_to(Ci[None, :], match_2d.shape)[match_2d]
+    s_idx_grid = np.broadcast_to(np.arange(len(Si))[:, None], match_2d.shape)
+    cs_s_indices_np = s_idx_grid[match_2d]
+    CSj = Sj[cs_s_indices_np]
+
+    Ii = Ij = np.arange(n_col)
+    I_CSi = jnp.asarray(np.concatenate([CSi, Ii], -1), dtype=jnp.int32)
+    I_CSj = jnp.asarray(np.concatenate([CSj, Ij], -1), dtype=jnp.int32)
+    cs_s_indices = jnp.asarray(cs_s_indices_np)
     symbolic = klujax.analyze(I_CSi, I_CSj, n_col)
 
     return (